CN117828216A - Multi-mode Web information retrieval static ordering learning method, system, equipment and medium - Google Patents

Abstract

The invention discloses a multi-mode Web information retrieval static ordering learning method, a system, equipment and a medium, wherein the method comprises the steps of obtaining webpage screenshot and webpage information of a target webpage and obtaining webpage text information, HTML text, HTML labels and webpage statistical characteristic information based on the screenshot and the information; mapping and extracting features of webpage text information, HTML tag information and webpage statistical feature information respectively to obtain text feature vectors, HTML tag feature vectors, HTML text feature vectors and webpage statistical feature vectors respectively; combining the characteristics of the obtained 4 characteristic vectors to obtain a combined characteristic vector; and performing evaluation and grading according to the combined feature vector to generate a predicted evaluation grade of the webpage. The invention explores the inherent characteristics of the webpage information deeply from multiple dimensions, evaluates the quality of the Web content of the target webpage more accurately, objectively and reasonably, improves the evaluation accuracy, and is not easily influenced by various cheating methods.

Description

Multimodal Web information retrieval static ranking learning method, system, device and medium

Technical Field

The present invention belongs to the technical field of web page information retrieval, and in particular relates to a multi-modal Web information retrieval static ranking learning method, system, device and medium.

Background technique

Web information refers to the large number of web pages and other related content on the Internet, including websites, blogs, social media posts, news articles, etc. Web information not only contains various high-quality information, but also contains various false, deceptive, cheating, and abused information. Therefore, in the context of a large amount of web content and mixed quality, how to identify high-quality web content has become increasingly important and urgent.

Static Ranking Learning for Web Information Retrieval, also known as Evaluation of Web Content Quality, Query Independent Ranking or Static Ranking, is used to evaluate the quality of Web content and classify or rank web content of different quality. Static Ranking Learning is the core and basic algorithm of search engines, recommendation systems, and intelligent dialogue services. For search engines, static ranking is the core algorithm of search engines. Its evaluation results serve as a reference for the "importance" of search rankings, and can also be used as an important basis for guiding the crawling of massive web pages.

At present, the static ranking of Web information retrieval has achieved some research results and applied various technologies, such as PageRank, TrustRank, Truncated PageRank and other link analysis algorithms. Link cheating has been well suppressed. However, there is a lack of effective solutions to deal with hidden cheating, embedded cheating, re-characterization cheating and Cloaking. At the same time, with the continuous diversification of Internet content forms and the continuous expansion of data scale, especially driven by interests, various abuses represented by Web cheating and false advertising are becoming increasingly rampant. How to efficiently identify Web content of different qualities is becoming more and more difficult. At the same time, netizens and large models such as LLM have an increasingly strong demand for high-quality information. At this time, the mainstream static ranking method of Web information retrieval is easily affected by abuse, resulting in less accurate ranking than expected, which brings new challenges to the existing static ranking method of Web information retrieval.

Summary of the invention

The main purpose of the present invention is to overcome the shortcomings of the prior art, that is, the low ranking accuracy and susceptibility to various cheating methods, and to propose a multimodal Web information retrieval static ranking learning method, system, device and medium.

In order to achieve the above object, the present invention adopts the following technical solutions:

A multimodal Web information retrieval static ranking learning method comprises the following steps:

Obtaining a webpage screenshot of the target webpage, obtaining webpage text information based on the webpage screenshot, and mapping the webpage text information into a first text vector sequence;

Obtaining HTML text of a target webpage, and mapping the HTML text information into a first HTML text vector sequence;

Obtaining HTML tags of the target web page, and mapping the HTML tag information into a first HTML tag vector sequence;

Based on the webpage statistical features of the target webpage, the webpage statistical feature information is mapped into a first webpage statistical feature vector sequence; the webpage statistical features include link analysis features, content heuristic features, website attribution features and time series features;

Based on the one-hot encoding and word embedding vector mode, a second text feature vector matrix, a second HTML text feature vector matrix and a second HTML tag feature vector matrix are obtained according to the first text vector sequence, the first HTML text vector sequence and the first HTML tag vector sequence; a Transformer encoder model is used to extract features from the second text feature vector matrix, the second HTML text feature vector matrix and the second HTML tag feature vector matrix to obtain a text feature vector, an HTML text feature vector and an HTML tag feature vector;

Based on the feature combination and concatenation model, the first webpage statistical feature vector sequence is concatenated to obtain a second webpage statistical feature vector matrix; the second webpage statistical feature vector matrix is subjected to feature extraction using a DNN model to obtain a webpage statistical feature vector;

Combining the text feature vector, the HTML text feature vector, the HTML tag feature vector and the webpage statistical feature vector to obtain a combined feature vector;

Evaluation and grading are performed based on the combined feature vector to generate a predicted evaluation grade for the web page.

The present invention also includes a multimodal Web information retrieval static ranking learning system, the system adopts the multimodal Web information retrieval static ranking learning method provided by the present invention, and the system includes an information extraction module, a feature extraction module, a feature combination module and a quality evaluation module;

An information extraction module is used to obtain a webpage screenshot and webpage information of a target webpage and obtain webpage text information, HTML text, HTML tags and webpage statistical feature information based on the webpage screenshot and webpage information;

A feature extraction module is used to map web page text information, HTML text information, HTML tag information and web page statistical feature information into a first text vector sequence, a first HTML text vector sequence, a first HTML tag vector sequence and a first web page statistical feature vector sequence, respectively; and is also used to perform feature extraction on the first text vector sequence, the first HTML text vector sequence, the first HTML tag vector sequence and the first web page statistical feature vector sequence, respectively, to obtain a text feature vector, an HTML text feature vector, an HTML tag feature vector and a web page statistical feature vector, respectively;

A feature combination module is used to combine the text feature vector, the HTML text feature vector, the HTML tag feature vector and the webpage statistical feature vector to obtain a combined feature vector;

A quality evaluation module is used to evaluate and grade the web page based on the combined feature vectors to generate a predicted evaluation grade for the web page;

The feature extraction module is equipped with a DNN model and a Transformer encoder model;

The feature combination module is provided with a fully connected layer, through which the text feature vector with text context feature information, the HTML text feature vector with HTML text context feature information, the HTML tag feature vector with HTML tag context feature information, and the web page statistical feature vector with web page statistical feature information are combined to calculate the best fusion method and obtain the feature results suitable for classification analysis;

The quality evaluation module is equipped with a DNN model, which evaluates and grades the quality of Web content based on the combined feature vector to determine the predicted evaluation grade of the current web page.

The present invention also includes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the multimodal Web information retrieval static ranking learning method provided by the present invention is implemented.

The present invention also includes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the multimodal Web information retrieval static ranking learning method provided by the present invention is implemented.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The present invention comprehensively considers multimodal factors. In addition to link analysis features, content heuristic features, web page attribution features and timing features, it further includes two different modal semantic information, namely text content information from a machine perspective, HTML tag information and web page screenshot OCR text information from a user perspective, and mines the above multi-perspective identification information through a deep learning point-wise architecture to achieve comprehensive consideration of multiple cheating factors such as link cheating, hidden cheating, embedded cheating, jump cheating, Cloaking, etc., while taking into account website content, attribution and timing characteristics; compared with the existing technology, it explores the intrinsic characteristics of web page information more deeply from multiple dimensions, evaluates the Web content quality of the target web page more accurately, objectively and reasonably, and improves the evaluation accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the method of the present invention;

Fig. 2 is a schematic diagram of the method of the present invention;

FIG3 is a schematic diagram of text feature vector transformation in an embodiment;

FIG4 is a schematic diagram of HTML text feature vector transformation in an embodiment;

FIG5 is a schematic diagram of the transformation of HTML tag feature vectors in an embodiment;

FIG6 is a schematic diagram of a web page statistical feature vector transformation in an embodiment;

FIG7 is a schematic diagram of a process of performing feature extraction using a Transformer encoder model in an embodiment;

FIG. 8 is a schematic diagram of the system of the present invention.

Detailed ways

The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

Example

As shown in FIG. 1 and FIG. 2 , the present invention, a multimodal Web information retrieval static ranking learning method, comprises the following steps:

Obtain a screenshot of the target webpage, obtain webpage text information based on the webpage screenshot, and map the webpage text information into a first text vector sequence, specifically:

For the acquired web page text information, the input data x _text is defined as a text sequence, and x _text will be divided into a series of discrete text units, that is, x _text = {t ₁ , t ₂ t ₃ , …, t _L }, where L is the original length of the input data, and _ti represents each discrete text unit in x _text ; by setting a fixed length as L _c , a new character element sequence is obtained after truncation or padding operation That is, the first text vector sequence, where/> Indicates/> Each discrete text unit in .

Get the HTML text of the target web page, and map the HTML text information into a first HTML text vector sequence, specifically:

The obtained HTML text information of the web page is defined as input data x _htmlText and regarded as an HTML text sequence. x _htmlText is divided into a series of discrete HTML text units, namely x _htmlText = {h ₁ ,h ₂ ,h ₃ ,…,h _hL }, where hL is the original length of the input data and _hi represents each discrete HTML text unit in x _htmlText . By setting a fixed length of L _h , a new HTML text element sequence is obtained after truncation or padding. That is, the first HTML text vector sequence, where /> Indicates/> Each discrete unit of HTML text in .

Get the HTML tag of the target web page, and map the HTML tag information into the first HTML tag vector sequence, specifically:

The obtained web page HTML tag information is defined as input data x _htmlTag and regarded as an HTML tag sequence. x _htmlTag will be divided into a series of discrete HTML tag units, that is, x _htmlTag = {g ₁ ,g ₂ ,g ₃ ,…,g _gL }, where gL is the original length of the input data and g _i represents each discrete HTML tag unit in x _htmlTag . By setting a fixed length of L _g , a new HTML tag element sequence is obtained after truncation or padding. That is, the first HTML tag vector sequence, where /> Indicates/> Each discrete HTML tag in a .

In this embodiment, obtaining a screenshot of the target webpage and obtaining webpage text information based on the screenshot includes:

Web page screenshots are collected based on link addresses (URLs); text box detection is performed on the web page screenshots to obtain all text box areas in the web page screenshots; an image recognition algorithm is used to perform text recognition on the web page screenshots according to the text box areas to obtain sequential web page text information.

In this embodiment, by making full use of the visual information in the web page, from the perspective of the web page user, using technologies such as OCR, various cheating methods can be circumvented to a certain extent, and real visible web page text information and web page image information can be directly extracted.

In some embodiments, a screenshot of the target web page is obtained through a strategy such as delayed waiting, and the screenshot is the image information required subsequently.

In some embodiments, for a web page screenshot, a web page text detection algorithm is used to detect all areas containing text in the web page screenshot and express them in the form of rectangular coordinates [x1, y1, x2, y2, x3, y3, x4, y4]. The horizontal and vertical coordinates of the four points at the lower left, upper left, lower right and upper right of the text box correspond to each other, and the matrix enclosed by them is called a text box.

In some embodiments, an OCR algorithm is used as the image recognition algorithm.

In some embodiments, an image processing function in the OpenCV function library is used as an image recognition algorithm.

In some embodiments, in order to ensure the accuracy of text recognition, text boxes with too small text box areas are filtered out, and then the relevant image processing functions in the OpenCV function library are used to extract the text box areas from the web page screenshot in sequence according to the text box coordinates and save them in the memory; if the text box area has an inclination angle, in order to improve the accuracy of text recognition, it is horizontalized and affine processed. After obtaining all the text box areas in the web page screenshot, multiple text box areas are extracted from the memory in parallel in the form of multi-threading, and the text box areas are input into the algorithm model (such as yolov4), and the corresponding text is identified according to the text features. After completing the text recognition of all text box areas, the identified texts are reassembled into complete web page texts in order to extract the text information of the web page and obtain the web page text information.

Based on the webpage statistical features of the target webpage, the webpage statistical feature information is mapped into a first webpage statistical feature vector sequence; the webpage statistical features include link analysis features, content heuristic features, website attribution features and time series features;

Based on the one-hot encoding and word embedding vector mode, a second text feature vector matrix, a second HTML text feature vector matrix and a second HTML tag feature vector matrix are obtained according to the first text vector sequence, the first HTML text vector sequence and the first HTML tag vector sequence; a Transformer encoder model is used to extract features from the second text feature vector matrix, the second HTML text feature vector matrix and the second HTML tag feature vector matrix to obtain a text feature vector, an HTML text feature vector and an HTML tag feature vector;

In this embodiment, as shown in FIG. 3 , FIG. 4 and FIG. 5 , specifically:

According to the unique hot encoding and word embedding vector mode, each discrete text unit Map it to a d-dimensional vector space representation/> Final first text vector sequence/> Mapping into vector matrix/> The matrix M _text is the second text feature vector matrix generated by feature construction; specifically, the value of d is 100;

According to the one-hot encoding and word embedding vector mode, each discrete HTML text unit Map it to a d-dimensional vector space representation/> Final first HTML text vector sequence/> Mapping into vector matrix/> The matrix M _htmlText is the second HTML text feature vector matrix generated by feature construction;

According to the one-hot encoding and word embedding vector mode, each discrete HTML tag unit Map it to a d-dimensional vector space representation/> Final first HTML tag vector sequence /> Mapping into vector matrix/> The matrix M _htmlTag is the second HTML tag feature vector matrix generated by feature construction.

As shown in FIG7 , the Transformer encoder model is used to extract features from the second text feature vector matrix, the second HTML text feature vector matrix, and the second HTML tag feature vector matrix, specifically:

The Transformer encoder model includes several encoders, which include a multi-head self-attention mechanism, a feedforward layer, and layer normalization;

The process of feature extraction using the Transformer encoder model is:

The feature vector matrix is positionally embedded to obtain a feature vector matrix with embedded position information. The Transformer encoder model cannot utilize the sequential information between words. In this embodiment, position embedding is introduced before text feature extraction to record the position information of words in the sequence, so that the Transformer encoder model can distinguish words in different positions, thereby acquiring the ability to capture sequential information.

The feature vector matrix of the embedded position information is passed through the multi-head self-attention mechanism, feedforward layer and layer normalization in each encoder in turn to finally obtain the feature vector;

For the feature extraction of the second text feature vector matrix, the process expression is:

Where _hi represents the vector matrix output by the i-th encoder, and l represents the number of encoders in the Transformer encoder model; A second text feature vector matrix representing embedded position information;

For the feature extraction of the second HTML text feature vector matrix, the process expression is:

in, A second HTML text feature vector matrix representing embedded position information;

For the feature extraction of the second HTML tag feature vector matrix, the process expression is:

in, A second HTML tag feature vector matrix representing the embedding position information;

Where _hi represents the vector matrix output by the i-th encoder, and l represents the number of encoders in the Transformer encoder model; Represents the word vector matrix;

The process of position embedding is as follows:

Wherein, pos represents the position number of the word, i represents the dimension number of the word, M _peText , M _peHtml and M _peHtmlTag represent the position information matrix of the second text feature vector matrix, the position information matrix of the second HTML text feature vector matrix and the position information matrix of the second HTML tag feature vector matrix respectively, and their dimensions are respectively equal to the dimensions of the second text feature vector matrix M _text , the second HTML text feature vector matrix M _html and the second HTML tag feature vector matrix; + represents a matrix addition operation;

The self-attention mechanism is the scaled dot product attention mechanism. The multi-head self-attention mechanism repeats the calculation process of the self-attention mechanism h times, and then concatenates the calculation results. The process expression of the multi-head self-attention mechanism is:

multiHead(Q,K,V)＝Concat(head ₁ ,…,head _h )

Among them, head _i represents the calculation result of the i-th self-attention mechanism, h represents the number of repetitions of the self-attention mechanism, and the value of h is 10; MultiHead(Q,K,V) represents the calculation result of the multi-head self-attention mechanism; Represents different linear transformation matrices; /> Represents query, key, and value respectively; d _k represents The number of columns of the matrix, i.e. the vector dimension; />

In any encoder, after the multi-head self-attention mechanism, the vector matrix A is obtained, and then input into the feedforward layer, which is composed of multiple fully connected layers and is responsible for performing vector matrix mapping and affine transformation on the vector matrix A; after residual connection and layer normalization, the output vector matrix h _i of the encoder module is obtained, and the output vector matrix continues to be used as the input of the next encoder module; the output h _l of the last encoder is the text feature vector or HTML text feature vector or HTML tag feature vector.

Based on the feature combination and concatenation model, the first webpage statistical feature vector sequence is concatenated to obtain the second webpage statistical feature vector matrix; for the webpage statistical feature information, it is defined as input data and regarded as a webpage statistical feature sequence, which is a one-dimensional numerical sequence, and the feature sequence also includes link analysis features x _link (TrustRank, SpamRank, Truncated PageRank, etc.), content heuristic features x _content (hot word coverage, text length, etc.), webpage attribution features x _attribution (domain name holder, IP address, etc.) and time series features x _time ;

As shown in FIG6 , the above four features are sequentially spliced and combined to generate a feature vector sequence x _statistics ={x _link ; x _content ; x _attribution ; x _time }; the feature vector sequence is dimensionally transformed to generate a feature vector matrix M _statistics ={M _link ; M _content ; M _attribution ; M _time };

Wherein M _statistics ∈R ^L*1 , L is the total number of web page statistical features, 1 represents that each element in the generated feature vector matrix is a 1-dimensional vector, and the M _statistics matrix is the second web page statistical feature vector matrix generated by feature construction.

Using a DNN model to extract features from the second webpage statistical feature vector matrix to obtain a webpage statistical feature vector; wherein the DNN model includes a plurality of fully connected networks;

The DNN model is used to extract features from the second webpage statistical feature vector matrix, specifically:

The second webpage statistical feature vector matrix is sequentially passed through each fully connected network for matrix operation, and finally the webpage statistical feature vector is obtained. The process expression is:

Among them, _Mi represents the vector matrix output by the i-th fully connected network, l represents the number of fully connected networks in the DNN model; M _statistics represents the statistical feature vector matrix of the second web page; the weight matrix of the fully connected network _Li is the number of neurons in the i-th fully connected network; _bi represents the bias term of the neurons in the i-th fully connected network; the output _Mi of the last fully connected network in the DNN model is the statistical feature vector of the web page.

The text feature vector, HTML text feature vector, HTML tag feature vector and web page statistical feature vector are combined to obtain a combined feature vector, specifically:

The text feature vector V _text , the HTML text feature vector V _htmlText , the HTML tag feature vector V _htmlTag and the webpage statistics feature vector V _statistics are sequentially concatenated and combined by using a vector concatenation operation to obtain a concatenated feature vector V _c =[V _text ; V _htmlText ; V _htmlTag ; V _statistics ];

A single fully connected layer is used to perform feature combination operation on the concatenated feature vectors. Its purpose is to remove redundant variables and find the intrinsic connection between feature vectors while performing feature dimensionality reduction through the fully connected layer. Its expression is:

V＝ _{WcVc + bc}

Among them, ^V∈Rf represents the feature combination vector, f represents the dimension of the low-dimensional feature space; the weight matrix _Wc∈Rf ^of the fully connected network, f is the number of neurons in the fully connected network; _bc represents the bias term of the neurons in the fully connected network.

Evaluation and grading are performed based on the combined feature vector V to generate a predicted evaluation grade for the web page, specifically:

The DNN model is used to evaluate and grade the combined feature vector V, where the DNN model is composed of multiple fully connected networks. Specifically, the combined feature vector V is used as the input of the DNN model, and matrix operations are performed in turn through multiple fully connected networks contained in the model. The final output is the predicted evaluation level of the combined feature vector, which is expressed as:

Among them, _Vi represents the vector matrix output by the i-th fully connected network, l _v represents the number of fully connected networks in the DNN model; the weight matrix of the fully connected network _fi is the number of neurons in the ith fully connected network; _bi is the bias term of the neurons in the ith fully connected network; the output of the last fully connected network in the DNN model/> That is the quality evaluation vector; /> Represents the predicted probability of each evaluation level corresponding to the quality evaluation vector;/> Indicates the predicted evaluation level;

In the iterative training process of the static ranking learning method for multimodal Web information retrieval, the multi-classification cross entropy function is used as the loss function. The weight parameters of all models in the back-propagation iterative update method are expressed as:

Among them, y ⁽ⁱ⁾ represents the true evaluation level of the i-th web page during the training process; It represents the predicted probability of each evaluation level corresponding to the i-th web page during the training process; N represents the total number of samples in the evaluation training process; i represents the index number of the current sample.

In some embodiments, y ⁽ⁱ⁾ ∈ [0, R], that is, y ⁽ⁱ⁾ is a quality assessment value in the interval [0, R]. In the present invention, referring to the ECML/PKDD Discovery Challenge rules and considering maximizing the quality difference as much as possible, the value of R is not less than 9, that is, the quality value is not less than 10 levels.

Those skilled in the art should understand that in the iterative training process of the multimodal Web information retrieval static ranking learning method, the parameters in all models will be back-propagated according to the difference between the predicted evaluation level and the actual evaluation level and the result of the loss function, and the loss function will be minimized using different algorithms such as gradient descent, and the parameters will be dynamically adjusted continuously to obtain the optimal parameters in continuous iterations.

In another embodiment, a multimodal Web information retrieval static ranking learning system is provided. The system adopts the multimodal Web information retrieval static ranking learning method of the above embodiment. As shown in FIG8 , the system includes an information extraction module, a feature extraction module, a feature combination module, and a quality evaluation module;

An information extraction module is used to obtain a webpage screenshot and webpage information of a target webpage and obtain webpage text information, HTML text, HTML tags and webpage statistical feature information based on the webpage screenshot and webpage information;

A feature extraction module is used to map web page text information, HTML text information, HTML tag information and web page statistical feature information into a first text vector sequence, a first HTML text vector sequence, a first HTML tag vector sequence and a first web page statistical feature vector sequence, respectively; and is also used to perform feature extraction on the first text vector sequence, the first HTML text vector sequence, the first HTML tag vector sequence and the first web page statistical feature vector sequence, respectively, to obtain a text feature vector, an HTML text feature vector, an HTML tag feature vector and a web page statistical feature vector, respectively;

A feature combination module is used to combine the text feature vector, the HTML text feature vector, the HTML tag feature vector and the webpage statistical feature vector to obtain a combined feature vector;

A quality evaluation module is used to evaluate and grade the web page based on the combined feature vectors to generate a predicted evaluation grade for the web page;

In this embodiment, the feature extraction module is provided with a DNN model and a Transformer encoder model;

In this embodiment, the feature combination module is provided with a fully connected layer, through which the text feature vector with text context feature information, the HTML text feature vector with HTML text context feature information, the HTML tag feature vector with HTML tag context feature information, and the web page statistical feature vector with web page statistical feature information are combined to calculate the best fusion mode and obtain the feature results suitable for classification analysis;

In this embodiment, the quality evaluation module is equipped with a DNN model, which evaluates and grades the quality of Web content based on the combined feature vector to determine the predicted evaluation grade of the current web page.

In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the multimodal Web information retrieval static ranking learning method as in the above embodiment is implemented.

The hardware entity of the computer device includes a processor, a memory and a communication interface; wherein the processor generally controls the overall operation of the computer device; the communication interface is used to enable the computer device to communicate with other terminals or servers through a network; the memory is configured to store instructions and applications executable by the processor, and can also cache data to be processed or processed by the processor and various modules in the computer device (including but not limited to image data, audio data, voice communication data and video communication data), which can be implemented through flash memory (FLASH) or random access memory (RAM).

Data can be transmitted between the processor, the communication interface and the memory via a bus, which may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together.

In another embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the multimodal Web information retrieval static ranking learning method of the above embodiment is implemented.

The storage medium may be transient or non-transient. Exemplarily, the storage medium includes, but is not limited to, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store computer program codes.

Exemplarily, the processor may be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA).

In another embodiment, a computer program is provided, including a computer-readable code. When the computer-readable code is executed in a computer device, a processor in the computer device executes part or all of the steps in the multimodal Web information retrieval static ranking learning method for implementing the above-mentioned embodiment.

In another embodiment, a computer program product is provided, which can be implemented by hardware, software or a combination thereof. As a non-limiting example, the computer program product can be embodied as a storage medium or a software product, such as an SDK (Software Development Kit).

It should also be noted that in this specification, terms such as "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A multi-mode Web information retrieval static ordering learning method is characterized by comprising the following steps:

acquiring a webpage screenshot of a target webpage, acquiring webpage text information based on the webpage screenshot, and mapping the webpage text information into a first text vector sequence;

acquiring an HTML text of a target webpage, and mapping the HTML text information into a first HTML text vector sequence;

acquiring an HTML tag of a target webpage, and mapping HTML tag information into a first HTML tag vector sequence;

mapping the webpage statistical feature information into a first webpage statistical feature vector sequence based on the webpage statistical features of the target webpage; the webpage statistical features comprise link analysis features, content heuristic features, website attribution features and time sequence features;

based on the single-hot coding and the word embedding vector mode, a second text feature vector matrix, a second HTML text feature vector matrix and a second HTML tag feature vector matrix are obtained according to the first text vector sequence, the first HTML text vector sequence and the first HTML tag vector sequence; performing feature extraction on the second text feature vector matrix, the second HTML text feature vector matrix and the second HTML tag feature vector matrix by adopting a transducer encoder model to obtain a text feature vector, an HTML text feature vector and an HTML tag feature vector;

vector stitching is carried out on the first webpage statistical feature vector sequence based on the feature combination and stitching model, and a second webpage statistical feature vector matrix is obtained; performing feature extraction on the second webpage statistical feature vector matrix by adopting a DNN model to obtain a webpage statistical feature vector;

combining the text feature vector, the HTML tag feature vector and the webpage statistical feature vector to obtain a combined feature vector;

and performing evaluation and grading according to the combined feature vector, and generating a predicted evaluation grade related to the webpage.

2. The method for learning static rank of multimodal Web information retrieval according to claim 1, wherein mapping the Web page text information into a first sequence of text vectors comprises:

for the acquired webpage text information, the input data x is defined _text And treat it as a text sequence, x _text Will be divided into a series of discrete text units, i.e. x _text ＝{t ₁ ,t ₂ ,t ₃ ,…,t _L Wherein L is the original length of the input data, t _i Represents x _text Each discrete text unit of (a); by setting a fixed length L _c Obtaining new character element sequence after cutting or filling operationI.e. a first text vector sequence, wherein +.>Representation->Each discrete text unit of (a);

mapping the HTML text information into a first HTML text vector sequence, specifically:

for the acquired webpage HTML text information, the input data x is defined _htmlText And treat it as a sequence of HTML text, x _htmlText Divided into a series of discrete HTML text units, i.e. x _htmlText ＝{h ₁ ,h ₂ ,h ₃ ,…,h _hL Where hL is the original length of the input data, h _i Represents x _htmlText Each discrete HTML text unit in (a); by setting a fixed length L _h Obtaining new HTML text element sequence after cutting or filling operationI.e. a first sequence of HTML text vectors, wherein +.>Representation->Each discrete HTML text unit in (a);

mapping the HTML tag information into a first HTML tag vector sequence, specifically:

for the obtained web page HTML tag information, it is defined as input data x _htmlTag And regards it as a sequence of HTML tags, x _htmlTag Will be divided into a series of discrete HTML tag units, i.e. x _htmlTag ＝{g ₁ ,g ₂ ,g ₃ ,…,g _gL gL is the original length of the input data, g _i Represents x _htmlTag Each discrete HTML tag unit of (a); by setting a fixed length L _g Obtaining new HTML tag element sequence after cutting or filling operationI.e. a first sequence of HTML tag vectors, wherein +.>Representation->Is provided.

3. The method for learning static ordering of multimodal Web information retrieval according to claim 2, wherein the second text feature vector matrix, the second HTML text feature vector matrix, and the second HTML tag feature vector matrix are obtained according to the first text vector sequence, the first HTML text vector sequence, and the first HTML tag vector sequence based on the one-hot encoding and the word embedding vector mode, specifically:

for each discrete text unit according to the one-hot encoding and word embedding vector modeMapping is performed, which is mapped to the vector space representation of d-dimension +.>Final first text vector sequence->Mapping into a vector matrixMatrix M _text Then a second text feature vector matrix generated by feature construction; d takes 100;

for each discrete HTML text unit, according to the one-hot encoding and word embedding vector patternsMapping is performed, which is mapped to the vector space representation of d-dimension +.>Final first HTML text vector sequence +.> Mapping into vector matrix>Matrix M _htmlText Then a second HTML text feature vector matrix generated by feature construction;

for each discrete HTML tag unit according to the one-hot encoding and word embedding vector modeMapping is performed, which is mapped to the vector space representation of d-dimension +.>Final first HTML tag vector sequence +.> Mapping into vector matrix>Matrix M _htmlTaG Then it is the second HTML tag feature vector matrix generated by the feature build.

4. The method according to claim 1, wherein the statistical feature information of the Web page is defined as input data and regarded as a statistical feature sequence of the Web page, the feature sequence is a one-dimensional numerical sequence, and the feature sequence includes link analysis features x _link Heuristic content feature x _content Web page attribution feature x _attribution And timing characteristics x _time ；

The 4 features are sequentially spliced and combined to generate a feature vector sequence x _statistics ＝{x _link ；x _content ；x _attribution ；x _time -a }; performing dimension conversion on the feature vector sequence to generate a feature vector matrix M _statistics ＝{M _link ；M _content ；M _attribution ；M _time }；

Wherein M is _statistics ∈R ^L*1 L is the total number of the statistical features of the webpage, 1 represents that each element in the generated feature vector matrix is a 1-dimensional vector, M _statistics The matrix is a second webpage statistical feature vector matrix generated by feature construction.

5. The method for learning static rank of multimodal Web information retrieval according to claim 1, wherein feature extraction is performed on the second text feature vector matrix, the second HTML text feature vector matrix, and the second HTML tag feature vector matrix by using a transducer encoder model, specifically:

the transducer encoder model comprises a plurality of encoders, wherein each encoder comprises a multi-head self-attention mechanism, a feedforward layer and layer normalization;

the process of feature extraction using a transducer encoder model is:

performing position embedding on the feature vector matrix to obtain a feature vector matrix embedded with position information;

normalizing the feature vector matrix embedded with the position information through a multi-head self-attention mechanism, a feedforward layer and a layer in each encoder in sequence to finally obtain feature vectors;

for feature extraction of the second text feature vector matrix, the process expression is as follows:

wherein h is _i A vector matrix representing the i-th encoder output, l representing the number of encoders in the transducer encoder model;a second text feature vector matrix representing embedded location information;

and extracting the characteristics of the second HTML text characteristic vector matrix, wherein the process expression is as follows:

wherein,a second HTML text feature vector matrix representing embedded location information;

and extracting the characteristics of the second HTML tag characteristic vector matrix, wherein the process expression is as follows:

wherein,representation inlayEntering a second HTML tag feature vector matrix of the position information;

the position embedding process specifically comprises the following steps:

wherein pos represents the position number of the word, i represents the dimension number of the word, M _peText 、M _peHtml M is as follows _peHtmlTag Respectively representing a position information matrix of a second text feature vector matrix, a position information matrix of a second HTML text feature vector matrix and a position information matrix of a second HTML tag feature vector matrix, the dimensions of which are respectively identical to those of the second text feature vector matrix M _text Second HTML text feature vector matrix M _html And the dimensions of the second HTML tag feature vector matrix are equal; + represents a matrix addition operation;

the self-attention mechanism is a zoom point multiplied attention mechanism, the multi-head self-attention mechanism is characterized in that the calculation process of the self-attention mechanism is repeated for h times, and then calculation results are spliced; the process expression of the multi-head self-attention mechanism is:

head _i ＝Attention(QW _i ^Q ,KW _i ^K ,VW _i ^V )

MultiHead(Q,K,V)＝Concat(head ₁ ,…,head _h )

wherein head _i Indicating the calculation result of the ith self-attention mechanism, wherein h indicates the repetition number of the self-attention mechanism, and the value of h is 10; multiHead (Q, K, V) represents the computation of a multi-headed self-attention mechanism; w (W) _i ^Q 、W _i ^K 、W _i ^V Representing different linear transformation matrices; QW (QW) _i ^Q 、KW _i ^K 、VW _i ^V Respectively representing a query, a key, and a value; d, d _k Representing QW _i ^Q 、KW _i ^K The column number of the matrix, i.e., the vector dimension;

in any encoder, after a multi-head self-attention mechanism, a vector matrix A is obtained, and then a feedforward layer is input, wherein the feedforward layer consists of a plurality of fully-connected layers and is responsible for mapping and affine transformation of the vector matrix A; after residual connection and layer normalization, an output vector matrix h of the encoder module is obtained _i And continuing the output vector matrix as input to the next encoder module; output h of last encoder _l Namely a text feature vector or an HTML tag feature vector.

6. The multi-modal Web information retrieval static ordering learning method according to claim 1, wherein the DNN model comprises a plurality of fully connected networks;

and adopting a DNN model to extract the characteristics of the second webpage statistical characteristic vector matrix, wherein the method specifically comprises the following steps:

and sequentially carrying out matrix operation on the second webpage statistical feature vector matrix in each fully-connected network to finally obtain the webpage statistical feature vector, wherein the process expression is as follows:

wherein M is _i Representing a vector matrix output by the ith fully-connected network, and l represents the number of fully-connected networks in the DNN model; m is M _statistics Representing a second web page statistical feature vector matrix; weight matrix of fully connected networkL _i Is the number of neurons in the ith fully-connected network; b _i A bias term representing a neuron in an ith fully-connected network; output M of last fully connected network in DNN model _i The statistical feature vector of the web page is obtained.

7. The method for learning static rank of multimodal Web information retrieval according to claim 1, wherein the feature combination of the text feature vector, the HTML tag feature vector and the Web page statistics feature vector is specifically:

using vector concatenation operation pairs to combine text feature vectors V _text HTML text feature vector V _htmlText HTML tag feature vector V _htmlTag Webpage statistics feature vector V _statistics Sequentially splicing and combining to obtain a spliced feature vector V _c ＝[V _text ；V _htmlText ；V _htmlTag ；V _statistics ]；

The method adopts a single full-connection layer to perform feature combination operation on spliced feature vectors, and aims to eliminate redundant variables and find out internal relations between the feature vectors while performing feature dimension reduction through the full-connection layer, wherein the expression is as follows:

V＝W _c V _c +b _c

wherein V is E R ^f Representing feature combination vectors, f representing the dimensions of the low-dimensional feature space; weight matrix W of fully connected network _c ∈R ^f F is the number of neurons in the fully connected network; b _c A bias term representing neurons in a fully connected network;

performing evaluation and grading according to the combined feature vector V to generate a predicted evaluation grade related to the webpage, wherein the predicted evaluation grade is specifically as follows:

the method comprises the steps of evaluating and grading a combined feature vector V by adopting a DNN model, wherein the DNN model consists of a plurality of fully-connected networks, specifically, taking the combined feature vector V as input of the DNN model, sequentially carrying out matrix operation through the plurality of fully-connected networks contained in the model, and finally outputting the predicted evaluation grade of the combined feature vector, wherein the expression is as follows:

wherein V is _i Vector matrix representing the output of the ith fully connected network, l _v Representing the number of fully connected networks in the DNN model; weight matrix of fully connected networkf _i Is the number of neurons in the ith fully-connected network; b _i A bias term representing a neuron in an ith fully-connected network; output of last fully connected network in DNN model +.>Namely, a quality evaluation vector; />Representing the prediction probability of the quality evaluation vector corresponding to each evaluation level; />Representing a predictive rating;

in the iterative training process of the multi-mode Web information retrieval static ordering learning method, a multi-classification cross entropy function is adopted as a loss function, and the weight parameters of all models in the iterative updating method are propagated in the opposite direction, wherein the expression of the loss function is as follows:

wherein y is ⁽ⁱ⁾ Representing the real evaluation grade of the ith webpage in the training process;representing the prediction probability of the ith webpage corresponding to each evaluation level in the training process; n represents the total number of samples in the evaluation training process; i represents the index number of the current sample;

in the iterative training process of the multi-mode Web information retrieval static ordering learning method, each parameter in all models is counter-propagated according to the difference between the predicted evaluation level and the real evaluation level and the result of the loss function, the loss function is minimized by using different algorithms such as gradient descent, and each parameter is dynamically adjusted continuously, so that the optimal parameter is obtained in continuous iteration.

8. A multi-mode Web information retrieval static ordering learning system, which is characterized in that the system adopts the multi-mode Web information retrieval static ordering learning method according to any one of claims 1-7, and comprises an information extraction module, a feature combination module and a quality evaluation module;

the information extraction module is used for acquiring webpage screenshot and webpage information of the target webpage and obtaining webpage text information, HTML text, HTML labels and webpage statistical characteristic information based on the webpage screenshot and the webpage information;

the feature extraction module is used for mapping the webpage text information, the HTML tag information and the webpage statistical feature information into a first text vector sequence, a first HTML tag vector sequence and a first webpage statistical feature vector sequence respectively; the method is also used for respectively carrying out feature extraction on the first text vector sequence, the first HTML tag vector sequence and the first webpage statistical feature vector sequence to respectively obtain a text feature vector, an HTML tag feature vector and a webpage statistical feature vector;

the feature combination module is used for carrying out feature combination on the text feature vector, the HTML tag feature vector and the webpage statistical feature vector to obtain a combined feature vector;

the quality evaluation module is used for performing evaluation and grading according to the combined feature vectors and generating a predicted evaluation grade related to the webpage;

the feature extraction module is provided with a DNN model and a transducer encoder model;

the feature combination module is provided with a full connection layer, and combines the text feature vector with text context feature information, the HTML text feature vector with HTML text context feature information, the HTML tag feature vector with HTML tag context feature information and the webpage statistics feature vector with webpage statistics feature information through the full connection layer, so as to calculate and obtain an optimal fusion mode, and obtain a feature result suitable for classification analysis;

the quality evaluation module is loaded with a DNN model, and evaluates and classifies the quality of the Web content based on the combined feature vector so as to judge the predicted evaluation grade of the current webpage.

9. A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor implements the multimodal Web information retrieval static ordering learning method of any of claims 1-7 when the computer program is executed.

10. A computer readable storage medium storing a computer program, characterized in that the multimodal Web information retrieval static ranking learning method of any one of claims 1-7 is implemented when the computer program is executed by a processor.